CN101697079B - Blind system fault detection and isolation method for real-time signal processing of spacecraft - Google Patents

Abstract

The invention provides a blind system fault detection and isolation method for real-time signal processing of a spacecraft, which comprises the following steps: according to object system diagnosis reference signals, adopting a time domain dynamic pattern matching mode to detect signal singular points to realize distinguishment of normal and abnormal states of a system; and then according to object system diagnosis reference signals after the fault occurrence time, adopting the time domain dynamic pattern matching mode to carry out matching and classification on the time domain signal pattern to realize system fault mode isolation. The blind system fault detection and isolation method is established on the basis of a fault detection ELMAN neural network, a fault isolation ELMAN neural network, an improved network training algorithm and fault logic judgment technology, has excellent real-time effectiveness, output coupling diagnosis performance, time domain signal diagnosis generalization and network convergence, and can effectively avoid the defects that the accurate spacecraft model is not easily acquired, an artificial diagnosis method has bad real-time, and the conventional neural network method has poor time domain sample generalization and convergence.

Description

The blind system fault that is used for the spacecraft real time signal processing detects and partition method

Technical field

The present invention relates to spacecraft fault diagnosis and fault-tolerant control field, particularly a kind of fault detect and partition method of handling in real time towards blind system time-domain signal.

Background technology

Real-time state monitoring and fault diagnosis are that the remote equipment reliability ensures requisite important means.Especially for spacecrafts such as satellite equipment, the key of the singularity of its remote environment and equipment determined the satellite failure diagnosis not only will possess intelligent independent, and need have real-time simultaneously.Infrared earth sensor is to be one of vitals of the reference measure attitude of satellite with orbit coordinate, also is the prerequisite that ensures the normal operation of satellite rail control subsystem.Yet sensor failure also is the easiest generation but one of the most scabrous control theory and technical matters.Be different from the actuator failures in the control system, sensor failure has propagated in control loop, generally can take two kinds of approach of model observer and neural network to carry out fault diagnosis.Model observer method has contained the Physical Mechanism of diagnosis object system, can generate the observation residual error to the defective space decoupling zero, carry out fault detect and isolation according to residual error, be fit to be applied to that mathematical description is simple in structure, the real-time monitoring and fault diagnosis of the accurate control system of modeling.The tradition neural net method because of its have function match and sort feature be applied to resolution system can't modeling or System Discrimination and pattern classification during more complicated.Can simulate the process model of objective system based on the System Discrimination of neural network, estimate the required identification variable of control system, yet, when the erratic complex fault pattern of diagnosis output, can detect but be not enough to isolated fault according to the System Discrimination variable; General what handle is not have correlativity, isolated data sample on the time based on the pattern classification of neural network, and needs abundant sample training could obtain stable network structure.For satellite attitude control system, owing to there is the complex uncertainty influence, set up comparatively difficulty of precise analytic model, and have nonlinear relationship between fault mode and the system's output.In addition, remote environment has determined the satellite failure diagnosis must have inline diagnosis ability faster.Thereby, study a kind of match that can make full use of neural network and classification capacity (not needing system is carried out modeling) but can realize the method for online real time fail diagnosis, for promoting satellite reliability and independence, have important theoretical and practice significance.

At present, the research of satellite sensor real-time fault diagnosis method mainly concentrates on model observer and two aspects of nerve network system identification.People such as Xing Yan have proposed a kind of observer method of utilizing feature structure to specify isolation satellite rolling driftage gyro sensor failure, utilize attitude motion to learn the system that relation is set up each sensor output information of related satellite, isolate observer at this system design, make the different directions orientation of the fault of different gyros in the residual error space, reach the purpose of fault isolation, yet this method is only at simple control system model, and it detects isolation adequate condition general being difficult in engineering and satisfies.Document " utilizing the new method of offering as a tribute observer diagnosis infrared earth sensor fault " has proposed a kind of based on the method for offering as a tribute observer diagnosis infrared earth sensor, by the incomplete subsystem of seeing is carried out the controllability decomposition, offer as a tribute observer to overcome the controllability condition restriction to seeing subsystem design again, can realize the detection and the isolation of infrared earth sensor, but this method at the be coupled measurement output (promptly having increased diagnosis basis information) of sensor gyro of system's output model, must carry out the residual error generation by resolving redundancy, thereby export when unreliable when gyro to measure, detect to isolate then and lost efficacy.People such as H.A.Taleb have proposed to detect the circulation neural net method of isolating satellite sensor and actuator failures based on the System Discrimination variable.The diagnosable simple satellite magnetic control system sensor fault of this method, owing to do not have coupled relation between the sensor output, thereby be easier to carry out fault isolation, but be not suitable for the relevant situation (as the infrared earth sensor fault in the complete satellite attitude control system) of sensor output under the fault condition according to the residual error ratio of Correlation Identification variable.Beijing Control Engineering Inst. has developed a cover satellite posture control system fault real-time simulation and a diagnosis algorithm based on the G2 expert system tool on embedded satellite appearance controlling fault analog platform, by the analysis of attitude observational variable being set up judgement priori rules knowledge base, can effectively detect isolation for specified fault.It is more loaded down with trivial details that yet G2 expert system rule knowledge base is manually set up process, and need with reference to a large amount of real-time status information in order to improve the fault detect isolation effect, thereby also have deficiency aspect intelligent.

Summary of the invention

The object of the present invention is to provide a kind of blind system fault that is used for the spacecraft real time signal processing to detect and partition method, this method has real-time effectiveness, output coupling diagnosis performance, time-domain signal diagnosis generalization and network convergence preferably.

The blind system fault that is used for the spacecraft real time signal processing detects and partition method, may further comprise the steps:

(1) online detection:

(1.1) with i=k-N ..., k detects spacecraft diagnosis basis live signal p constantly _iInput fault detects neural network a ₂(i)=f ₂(LW _2,1a ₁(i)+b ₂), a wherein ₁(i)=f ₁(IW _1,1p _i+ LW _1,1a ₁(i-1)+b ₁), IW _1,1, LW _1,1, b ₁, LW _2,1, b ₂Be the structural parameters of fault detect neural network, f ₁() is the hidden layer transition function of fault detect neural network, f ₂() is the output layer transition function of fault detect neural network, a ₂(i) be one-dimensional vector, a ₁(i) be the h dimensional vector, h 〉=m (m*1)/2, m are diagnosis basis live signal p _iNumber, k be current detection constantly, N is a data window length;

(1.2) to i=k-N ..., k detects fault detect neural network output a constantly ₂(i) make Filtering Processing, obtain the filtering result $O\left(k\right)=\frac{1}{N+1}\underset{i=k-N}{\overset{i=k}{\mathrm{\Σ}}}{a_2}^2\left(i\right);$

(1.3) if filtering as a result O (k) smaller or equal to detection threshold ε ₀, then system is normal, otherwise the system failure enters step (2);

(2) online isolation:

(2.1) with i=k-N ..., k detects spacecraft diagnosis basis live signal p constantly _iInput fault is isolated neural network ${\tilde{a}}_2\left(i\right)={\tilde{f}}_2(\tilde{L}{W}_{\mathrm{2,1}}{\tilde{a}}_1\left(i\right)+{\tilde{b}}_2),$ Wherein ${\tilde{a}}_1\left(i\right)={\tilde{f}}_1(\tilde{I}{\mathrm{<figure-callout\; id="1"\; label="W"\; filenames="H2009102722656E00021.png,H2009102722656E00033.png"\; state="\{\{state\}\}">W</figure-callout>}}_{\mathrm{1,1}}{p}_i+\tilde{L}{\mathrm{<figure-callout\; id="1"\; label="W"\; filenames="H2009102722656E00021.png,H2009102722656E00033.png"\; state="\{\{state\}\}">W</figure-callout>}}_{\mathrm{1,1}}{\tilde{a}}_1(i-1)+{\tilde{b}}_1),$

Be the structural parameters of fault isolation neural network,

Be the hidden layer transition function of fault isolation neural network,

Be the output layer transition function of fault isolation neural network,

For

Dimensional vector, $\tilde{n}\&GreaterEqual;{\log }_2\mathrm{FN},$ FN is the fault sum,

For

Dimensional vector, $\tilde{h}\&GreaterEqual;(m*\tilde{n})/2;$

(2.2) to i=k-N ..., k detects fault isolation neural network output constantly

Make Filtering Processing, obtain the filtering result $\tilde{O}\left(k\right)={\left[\begin{array}{ccc}{\tilde{O}}_1& ...& {\tilde{O}}_{\tilde{n}}\end{array}\right]}^T=\frac{1}{N+1}\underset{i=k-N}{\overset{i=k}{\mathrm{\&Sigma;}}}{{\tilde{a}}_2}^2\left(i\right),$ T represents transposition;

(2.3) the filtering result that step (2.2) is obtained

In each the dimension element

Respectively with isolation threshold

Make comparisons, smaller or equal to zero, then the logic output valve of this dimension element correspondence is H0 as if comparative result, otherwise the logic output valve of this dimension element correspondence is H1, thereby obtains the filtering result vector Corresponding logic output sequence;

(2.4) in the fault mode matching list, mate fault mode according to the logic output sequence.

The structural parameters IW of described fault detect neural network _1,1, LW _2,1, b ₁, LW _2,1, b ₂Determine in the following manner:

A1, make X (t)=[IW _1,1(t) LW _2,1(t) b ₁(t) LW _2,1(t) b ₂(t)] ^T, initialization X (1), X (2), t=3;

A2, calculating $\left\{\begin{array}{c}X\left(t\right)=X(t-1)+\mathrm{\&Delta;X}(t-1)\\ \mathrm{\&Delta;X}(t-1)=\mathrm{\&alpha;\&Delta;X}(t-2)+\mathrm{\&alpha;}{l}_r\frac{d_{\mathrm{perf}}}{\mathrm{dX}}\end{array}\right.,$ Wherein, detect neural network performance perf and the derivative that detects weight and migration parameter X (t) $\frac{d_{\mathrm{perf}}}{\mathrm{dX}\left(t\right)}=\frac{d}{\mathrm{dX}\left(t\right)}\left(\underset{s=1}{\overset{G}{\mathrm{\&Sigma;}}}{(a_2\left(s\right)-y\left(s\right))\&CenterDot;(a_2\left(s\right)-y\left(s\right))}^T\right),$ Y (s) is the desired output of s fault detect neural network during the moment, and G is the time series total length of normal and fault training sample, l _rThe learning rate of expression fault detect neural network, α represents the momentum term of fault detect neural network;

A3, if detect neural network performance perf ∈ [0,0.001], the structural parameters equivalence of then fault detect neural network finish, otherwise t=t+1 returns steps A 2 in X (t).

The structural parameters of described fault isolation neural network Determine in the following manner:

B1, order $\tilde{X}\left(t\right)={\left[\begin{array}{ccccc}\tilde{I}{\mathrm{<figure-callout\; id="1"\; label="W"\; filenames="H2009102722656E00021.png,H2009102722656E00033.png"\; state="\{\{state\}\}">W</figure-callout>}}_{\mathrm{1,1}}\left(t\right)& \tilde{L}{W}_{\mathrm{2,1}}\left(t\right)& {\tilde{b}}_1\left(t\right)& \tilde{L}{W}_{\mathrm{2,1}}\left(t\right)& {\tilde{b}}_2\left(t\right)\end{array}\right]}^T,$ Initialization T=3;

B2, calculating $\left\{\begin{array}{c}\tilde{X}\left(t\right)=\tilde{X}(t-1)+\mathrm{\&Delta;}\tilde{X}(t-1)\\ \mathrm{\&Delta;}\tilde{X}(t-1)=g\left(\mathrm{\&Delta;}\tilde{X}(t-2)\right)\mathrm{\&Delta;}\tilde{X}(t-2)+\widetilde{\mathrm{\&alpha;}}{\tilde{l}}_r\frac{d_{p\tilde{e}\mathrm{rf}}}{d\tilde{X}\left(t\right)}\end{array}\right.,$ Wherein, isolation network performance

With isolation weight and migration parameter

Derivative $\frac{d_{p\tilde{e}\mathrm{rf}}}{d\tilde{X}\left(t\right)}=\frac{d}{d\tilde{X}\left(t\right)}(\underset{\tilde{s}=1}{\overset{\tilde{G}}{\mathrm{\&Sigma;}}}({\tilde{a}}_2\left(\tilde{s}\right)-\tilde{y}\left(\tilde{s}\right))\&CenterDot;{({\tilde{a}}_2\left(\tilde{s}\right)-\tilde{y}\left(\tilde{s}\right))}^T),$

For

The desired output of fault isolation neural network in the time of constantly,

Be the length of time series of fault training sample,

The learning rate of expression fault isolation neural network,

The momentum term of expression fault isolation neural network, $g\left(\mathrm{\&Delta;}\tilde{X}(t-2)\right)=\mathrm{\&Delta;}\tilde{X}(t-2)\&CenterDot;\exp (-\mathrm{\&Delta;}{\tilde{X}}^2(t-2));$

B3, if isolate the neural network performance $p\tilde{e}\mathrm{rf}\&Element;\left[\mathrm{0,0.001}\right],$ Then the structural parameters equivalence of fault isolation neural network in

Finish, otherwise t=t+1 returns step B2.

Described transition function f ₁(), f ₂(),

In employing tansig (), logsig (), the purelin () function any one.

Data window length N value hour helps the fast detecting to fault, yet might increase rate of false alarm; Otherwise, when the N value is excessive, will be unfavorable for fast detecting, general value 1～20 to fault.

Technique effect of the present invention is embodied in: the present invention has substantive distinguishing features and marked improvement, and it is to research and develop on the basis of the good fault detect ELMAN neural network of dynamic property, fault isolation ELMAN neural network, improvement network training algorithm, fault logic Technology of Judgment that blind system fault detects with partition method.Compare with existing technology, this technology has real-time effectiveness, output coupling diagnosis performance, time-domain signal diagnosis generalization and network convergence preferably, can avoid that the spacecraft valid model be difficult for to obtain, the defective of artificial diagnostic method real-time poor, traditional neural net method time domain samples generalization and poor astringency, the automatic fault diagnosis ability and the reliability of blind systems such as spacecraft improved and improves.

Description of drawings

Fig. 1 is principle of the invention figure;

Fig. 2 is a system architecture synoptic diagram of the present invention;

Fig. 3 is an ELMAN neural network structure synoptic diagram;

Fig. 4 is a satellite posture control system absolute orientation complete model synoptic diagram;

Fig. 5 is the interface relationship synoptic diagram of real-time diagnosis module and objective system;

Fig. 6 shuts fault detect isolation effect figure for roll channel output;

Fig. 7 is roll channel output mean bias fault detect isolation effect figure;

Fig. 8 shuts fault detect isolation effect figure for jaw channel output;

Fig. 9 is jaw channel output mean bias fault detect isolation effect figure;

Figure 10 is a fault infrared earth instrument coupling output signal synoptic diagram;

Figure 11 is fault isolation ELMAN neural network diagnosis output effect figure;

Figure 12 is diagnosis live signal synoptic diagram;

Figure 13 is BPNN and ELMAN NN of the present invention diagnosis output contrast synoptic diagram;

Figure 14 is not for improving ELMAN NN training constringency performance synoptic diagram;

Figure 15 improves ELMAN NN training constringency performance synoptic diagram for the present invention.

Embodiment

As shown in Figure 1, the present invention is divided into off-line and online two stages.Off-line phase: the first step, diagnosis basis signal number (comprise the control input and measure two types of outputs) and needs that can be for reference according to objective system detect the basic structure that the number of defects of isolating is determined fault detect and two kinds of ELMAN neural networks of fault isolation; Second step, sample data and the setting training objective of intelligence-collecting object system under normal and fault mode, adopt improved renewal gradient policy to carry out structure and weight parameter that off-line training obtains two kinds of networks respectively, and then obtain optimum fault detecting neural network and fault isolation neural network module; The 3rd step, the corresponding fault logic judging module of design after two kinds of neural network module output.The online stage: two kinds of networks and corresponding judging module embedded object system are carried out on-line real time monitoring, startup when the fault isolation neural network produces command signal (detecting fault takes place) in the judgement of fault detect neural network counterlogic, and then obtain the corresponding accurately command signal of fault isolation; In conjunction with fault detect command signal and fault isolation commands signal, can realize the real time fail diagnosis.

The present invention has fault detect and isolates neural network (comprising the off-line training algorithm), fault detection logic judgement, fault isolation neural network (comprising the off-line training algorithm), four parts of fault isolation logical decision and function (as shown in Figure 2).Wherein, fault detect is isolated the neural network module function for adopting the mode of time domain dynamic mode coupling to realize the detection of Singular Point according to objective system diagnosis reference signal, and then the differentiation of realization system's normal condition and abnomal condition, its off-line training at sample of signal be a time series process, training objective is normal and improper two states.Fault isolation neural network module function is for adopting the mode of time domain dynamic mode coupling to realize the coupling and the classification of time-domain signal pattern according to the diagnosis of the objective system after fault generation moment point reference signal, and then the isolation between the realization system failure pattern, its off-line training at sample of signal be a time series process, training objective is the scale-of-two sequence number collection of fault mode correspondence.Fault detect and fault isolation logical decision functions of modules generate accurate command signal for adopting the voting logic decision method.

Since at be that a class time domain Dynamic Signal carries out real time fail and detects and isolate, two neural networks of the present invention all adopt Ai Man (ELMAN) neural network with time delay memory characteristic as architecture prototyping.The ELMAN network has the sandwich construction similar to the multilayer feedforward network, as shown in Figure 3, its primary structure is that feedforward connects, comprise input layer (input layer), hidden layer (recurrentlayer), output layer (output layer), six structural parameters are learnt to revise by the off-line training algorithm; Feedback connects by one group of " structure " unit formation, is used for remembering the output valve of previous moment.

The present invention is described in more detail below in conjunction with embodiment and Figure of description:

In order to verify the validity and the superiority of the ELMAN neural network real-time fault diagnosis method that invention proposes, this paper is experimental subjects with certain type high precision three-axis stabilization near-earth satellite posture control system absolute orientation complete model, designed the real-time simulation experiment from real-time effectiveness, output coupling diagnosis performance, three aspects of time-domain signal diagnosis generalization contrast respectively, this experiment is finished based on Beijing Control Engineering Inst.'s embedded type fault analog platform.

One, objective system is described

1, complete satellite attitude control system model

Certain type high precision three-axis stabilization near-earth satellite posture control system absolute orientation complete model pie graph as shown in Figure 4, its topworks is a counteraction flyback, its attitude sensor comprises the responsive instrument of rate integrating gyroscope and infrared earth.Whole satellite posture control system is made up of controller, topworks, attitude of satellite kinetic model, Satellite Attitude Movement model, attitude sensor and attitude determination module.

2, infrared earth instrument fault

Usually comprise in the satellite posture control system attitude sensor and roll and two passage infrared earth sensors of driftage, the yaw-position angle, roll attitude angle of instrumented satellite constitutes Satellite Attitude Determination System instrumented satellite attitude with inertia gyroscope respectively.The infrared earth instrument comprises normal mode and two kinds of basic fault patterns, i.e. output is shut, the output mean bias.

Under the normal condition, the measurement model of infrared earth sensor can be abbreviated as:

$\left\{\begin{array}{c}{\mathrm{\&phi;}}_g=\mathrm{\&phi;}+{\mathrm{\&phi;}}_e\\ {\mathrm{\&theta;}}_g=\mathrm{\&theta;}+{\mathrm{\&theta;}}_e\end{array}\right.$

φ wherein _g, θ _gFor measuring output, φ, θ are that satellite rolls and the pitch attitude angle φ _e, θ _eFor measuring noise.

When hot outer terrestrial globe existed output to shut fault, its phenomenon of the failure was that the insensitive attitude of satellite of infrared earth sensor changes, and output keeps a solid stable constant value or other random variation amounts, and corresponding fault model mathematical description is:

$\left\{\begin{array}{c}{\mathrm{\&phi;}}_g={\mathrm{\&phi;}}_f\\ {\mathrm{\&theta;}}_g={\mathrm{\&theta;}}_f\end{array}\right.$

φ wherein _f, θ _fOutput valve when being fault is generally a normal value.

When there was the output mean bias in hot outer terrestrial globe, its phenomenon of the failure was that infrared measurement output average has deviation, departs from normal value, and corresponding fault model mathematical description is:

$\left\{\begin{array}{c}{\mathrm{\&phi;}}_g=\mathrm{\&phi;}+b_{\mathrm{\&phi;}}+{\mathrm{\&phi;}}_e\\ {\mathrm{\&theta;}}_g=\mathrm{\&theta;}+b_{\mathrm{\&theta;}}+{\mathrm{\&theta;}}_e\end{array}\right.$

Two, operational example flow process

According to aforesaid operations example objective for implementation, determine that objective for implementation diagnosis basis live signal is 6 the tunnel, i.e. 4 road actuator control moments input, the infrared earth instrument rolls, driftage two-way sensor is measured output.The interface relationship of real-time diagnosis module and objective system as shown in Figure 5.The number of defects that needs to detect with isolating is 4.The operational example flow process is divided into: sample data is obtained, network basic structure is determined, network training and parameter is determined, online fault detect with isolate four steps.

2.1, sample data obtains

Obtain 6 road live signal data samples under fault and the normal mode respectively, wherein fault mode comprises four kinds of concrete fault modes, for the discrete data that obtains conforming with stationarity and normality, being fit to Computer Processing, need the pre-service of dispersing of signal data sample, can adopt the extraction trend term, eliminate noise, reject wild point, prior aries such as normalized.This 6 road live signal is: the infrared earth instrument rolls, driftage two-way sensor measures output and 4 road actuator control moments are imported,

2.2, network basic structure determines

2.2.1 fault detect ELMAN neural network

For fault detect ELMAN neural network, be input as the m=6 road, owing to only need normal and improper the differentiation, thereby be output as the n=1 road.Thereby can get fault detect hidden layer neuron number: h 〉=(m*n)/2=1 * 6/2=3, select h=10; Wherein input layer is chosen as the tansig function to the neuron transition function of hidden layer, and hidden layer is chosen as the purelin function to the neuron transition function of output layer.

Training objective

Fault detect ELMAN neural metwork training target can be expressed as shown in the table 1, and target is output as 0 under the normal mode, is output as 1 under the fault mode.

Input of table 1 fault detect ELMAN network training and output corresponding tables

Input pattern	Normally	Fault is shut in roll channel output	Roll channel output mean bias fault	Fault is shut in jaw channel output	Jaw channel output mean bias fault
						Target output	H0	H1	H1	H1	H1

2.2.2 fault isolation ELMAN neural network

For fault isolation ELMAN neural network, be input as the m=6 road equally, owing to need two passages, two class faults, 4 kinds of faults are altogether distinguished, thereby be output as $\tilde{n}={\log }_{2}4=2$ The road.Thereby can get fault isolation hidden layer neuron number: $\tilde{h}=(m*\tilde{n})/2=2\&times;6/2=6,$ Select $\tilde{h}=10;$ Wherein input layer is chosen as the tansig function to the neuron transition function of hidden layer, and hidden layer is chosen as the tansig function to the neuron transition function of output layer.

Training objective

ELMAN neural metwork training target can be expressed as shown in the table 2.

Input of table 2 fault isolation ELMAN network training and output corresponding tables

Input pattern	Fault is shut in roll channel output	Roll channel output mean bias fault	Fault is shut in jaw channel output	Jaw channel output mean bias fault
					Target output	[H0?H0]	[H0?H1]	[H1?H0]	[H1?H1]

2.3 network training and parameter are determined

2.3.1 fault detect ELMAN neural network parameter is selected:

Because fault detect ELMAN neural network has only 1 tunnel output and pattern simpler, adopts gradient descent algorithm (traingdx) algorithm to carry out network training, normal and fault sample is as training sample with all, its selection learning rate l _r=0.8, momentum term α=0.5, network weight and threshold parameter IW _1,1, LW _1,1, b ₁, LW _2,1, b ₂Initial value is set to 0; Data window length N=10.

It is as shown in the table to obtain the optimal network parameter by network training:

IW _1,1Be the matrix of 6*15, its value is:

-14.4088?3.941136?3.4694?4.645904?11.52946?-22.2844?28.98437?-18.0139?26.98045?-12.1813?13.99182 5.6736?-11.4558?31.69466?136.128 -2.24172?3.049802?2.052508?3.909641?4.192213?-3.46149?5.141702?-4.98175?4.099?-3.39223?-8.18644?-4.36636?-2.27838?15.87086?44.29894 -17.5488?-12.075?1.0983991?5.41916?7.644464?-5.15613?6.0015791?4.28974?-8.00611?-8.35153?-1.9519?8.511803?-13.7594?5.049633?-5.66265 -19.5256?22.4005?17.40089?-20.3509?10.98958?-14.7693?-21.0067?-14.8935?27.18857?-16.2751?21.92103?14.94894?22.91684?-37.0329?-43.8502 -5.6343?-4.11538?-15.6593?20.70659?-20.196?9.599917?-14.6369?-21.7877?9.350528?27.06233?14.68206?-23.4725?-18.4324?-21.6379?-53.3853 5.16811?0.547195?-12.1259?2.816964?-8.30533?-11.8289?-3.39077?-7.21369?-0.30988?-10.1559?-14.0563?-6.82597?-10.1308?7.212921?9.78369

LW _1,1Be the matrix of 15*15, its value is:

-0.16449?0.170522?-0.1407?0.052302?-0.37242?0.199613?-0.0322?-0.07228?0.506883?-0.4678?0.2396?-0.55562?0.103322?-0.14302?-0.18192 0.23405?0.210114?0.358946?0.099499?-0.04655?-0.38066?-0.35991?0.443438?-0.13074?-0.10527?0.000473?0.069311?-0.09331?0.333803?-0.10561 -0.30216?0.160331?0.524273?-0.35125?0.304306?-0.13351?0.001982?0.485634?-0.17149?0.167222?-0.12225?-0.40509?0.463158?0.442383?0.011738 -0.29242?0.453838?-0.36419?0.20104?-0.03072?0.41907?-0.51079?-0.05712?-0.14716?-0.54243?0.524062?0.585757?-0.33298?-0.14511?0.205949 -0.209?0.458059?-0.37978?0.248813?-0.41016?-0.22963?0.312004?-0.41771?-0.3364?0.318894?-0.24654?0.173834?0.286702?0.239292?0.213027 -0.1674?0.262659?0.339014?0.207499?-0.32603?-0.23852?-0.20023?0.585398?0.518588?0.239998?0.411141?-0.16989?0.416787?0.303115?0.190125 -0.50495?0.530367?0.336873?-0.20212?-0.47333?-0.04652?0.078422?0.031973?-0.44404?0.370744?0.313655?0.263442?0.40584?-0.18976?-0.01476 0.517766?0.472448?-0.42286?0.579437?0.131483?-0.51133?-0.05834?-0.22786?-0.0931?-0.27253?-0.32133?0.577621?0.430129?0.43433?0.550961 0.487925?-0.46448?0.496241?0.431749?-0.50618?-0.31695?0.462823?-0.04679?-0.09482?-0.51297?0.003581?-0.15076?-0.1907?-0.56085?0.348278 0.506721?-0.14274?-0.0793?0.434471?-0.42381?0.561276?0.025689?-0.20241?-0.01632 0.2504?-0.38023?0.079008?0.148771?-0.39546?0.584359 -0.31222?0.399599?-0.32562?-0.0484?0.41899?0.105791?-0.49533?-0.26154?-0.45115?-0.34138?0.216693?0.067418?-0.37573?0.552533?-0.60885 0.162091?-0.42763?-0.14267?-0.20578?-0.42488?-0.28054?0.428221?0.207993?-0.00327?-0.04699?0.062469?-0.21411?-0.39629?-0.43743?0.110906 -0.14603?-0.45088?-0.20615?0.110259?0.001906?0.344175?0.348098?0.466686?0.126259?-0.55072?0.56284?0.50503?-0.4395?0.18888?-0.57112 0.396196?0.072512?-0.23004?-0.46461?-0.46256?0.575139?-0.43682?0.18317?-0.23441?0.070013?0.429853?0.374348?0.282232?0.080984?-0.14832 0.406014?0.261764?-0.52618?-0.53739?0.36405?0.106421?0.173308?0.103365?-0.43981?0.216083?0.486268?-0.15031?-0.23471?-0.38902?0.220685

b ₁Be the 15*1 matrix, its value is

-7.05542 -5.03347
	-10.9287 -5.67354 4.495754 3.814751 4.752147 -7.03463 6.911845 -4.66504 -4.42513 4.997596 -4.6552 4.926692 -2.83256

LW _2,1Be the 15*1 matrix, its value is

16.79143 -7.04655 -6.64024 -5.74963 -3.1712 6.27725 13.16769 13.23661 -8.93964 -1.03838 -2.89703 4.684465 -3.80658 32.13311 50.80333

b ₂Be the 1*1 matrix, its value is:

7.9346

2.3.2 fault isolation ELMAN neural network parameter is selected

As training sample, select learning rate with fault sample ${\tilde{l}}_r=0.8,$ Momentum term $\widetilde{\mathrm{\&alpha;}}=0.5,$ Network weight and threshold parameter

Initial value is set to 0; Data window length N=10.

It is as shown in the table that training obtains the optimal network parameter:

Be the matrix of 6*15, its value is:

33.34758?15.32701?-6.41953?-39.2418?36.81608?47.79748?-36.5888?34.83486?48.91323?-35.6004?-2.92682 35.3333?30.57198?30.20434?7.021055 3.759792?-5.83246?1.495785 -15.677?8.546199?8.202857?9.150806?-12.9321?15.5766?-9.75467?13.03573?1.022959?-17.1703?-9.87885?4.705955 33.51339?-39.448?-26.603?-20.4763?18.71537?17.5182?6.380108?26.80078?-3.71062?27.5587?-22.1752?18.02801?8.367109?-13.695?-29.8713 63.71129?60.90172?-79.3172?-22.7928?63.1518?67.64109?75.75815?-11.9603?-12.8395?-21.8333?-9.48426?59.94551?50.19797?70.63135?-92.4378 -44.2167?41.92115?-64.4447?-31.4492?65.2174?-42.4983?-26.3784?-6.25366?-29.0774?-56.9869?1.1883?-51.1642?13.58594?-69.2699?36.10263 -11.7007?12.26223?-16.1732?-16.2357?-6.85559?3.354294?-25.0714?-25.2094?-22.9134?-21.6735?-37.9796?-30.9255?15.98611?2.733477?2.98067

Be the 15*15 null matrix.

Be the 15*1 matrix, its value is:

-2.43367 0.710791 -4.97223 1.517838 2.158751 -1.93458 2.55884 -0.93891 -3.50157 -0.89673 -2.00239 0.43693 3.987831 2.474754 -0.45203

Be the 15*2 matrix, its value is:

-0.27437?2.108781 0.201598?-7.10501 -0.12267?0.863899 0.104398?1.711708 -0.12821?-3.95926 -0.54018?2.803249 0.301377?-1.91677 0.6091?-2.755 0.009536?-2.67891 -0.03464?-4.70647 -0.31738?-0.72685 0.73315?3.713391 0.028693?-0.57466 -0.24881?-2.04252
	0.728226?-2.43094

Be the 1*2 matrix, its value is:

-0.74981 -0.35202

2.4, the online in real time fault diagnosis

Six tunnel real-time diagnosis signals input fault simultaneously detect ELMAN neural network and fault isolation ELMAN neural network, there is the influence of uncertain shake in output in view of neural network, and the output of two neural network algorithm modules is all handled by the fault judging module.After first judging module has detected the fault generation, starting fault isolation ELMAN neural network handles the real-time diagnosis signal, stop fault detect ELMAN neural network simultaneously and proceed fault detect, after treating that second judging module is handled acquisition fault isolation result, can produce accurate fault diagnosis result signal in conjunction with the detection information of the first fault judging module.

For first judging module, if filtering as a result O (k) smaller or equal to detection threshold ε ₀, then H0 sets up, otherwise the H1 establishment, its decision rule is as follows:

If H0 sets up, then non-fault takes place, and then is left intact, and the judgement non-fault takes place;

If H1 sets up, then there is fault to take place, start the fault isolation module, stop fault detection module;

Detection threshold ε ₀General span is 0～0.5, and value 0.15 in the example.

For second judging module, it is output as 2 dimensional signals, with the filtering result

In the bidimensional element respectively with isolation threshold

Make comparisons, smaller or equal to zero, then H0 sets up as if comparative result, otherwise the H1 establishment, thereby obtain

The logic output sequence, in the fault mode matching list, mate fault mode according to the logic output sequence, the fault mode matching list is table 2, and is specific as follows:

If output [H0 H0] is set up, then adjudicate roll channel output and shut the fault generation;

If output [H0 H1] is set up, then adjudicate roll channel output mean bias fault and take place;

If output [H1 H0] is set up, then adjudicate jaw channel output and shut the fault generation;

If output [H1 H1] is set up, then adjudicate jaw channel output mean bias fault and take place;

Otherwise annulment.

Isolation threshold

General span is 0～0.5, and value 0.15 in the example.

Three, operational example effect

A. real-time effectiveness

Set respectively that four kinds of infrared earth instrument faults all take place at moment 1500s in the foregoing description, then the real time fail of four kinds of fault modes detects and isolates curve shown in Fig. 6～9. wherein, observing fault detect ELMAN neural network curve of output (NN1Out) is not difficult to find out, in very short time afterwards takes place in the 1500s fault, fault detect ELMAN neural network curve of output has original zero-mean to depart from detection threshold gradually, and fault detect comes into force.Detect after the fault generation, enable fault isolation ELMAN neural network and diagnose.Continuing to observe fault isolation ELMAN neural network output result (NN2Out1 and NN2Out2) curve is not difficult to find out, when fault causes corresponding output to depart from the saltus step exceeded threshold gradually by original 0 average, neural network output still can be stabilized in corresponding threshold range, keeps the validity of diagnostic result.

B. export the coupling diagnosis performance

In order to prove the validity of ELMAN neural network under the output signal coupling condition, the fault diagnosis result that this paper selects the output of infrared earth instrument roll channel to shut under the fault mode is verified.Observe neural network input signal curve (Figure 10) and diagnose curve (Figure 11) as can be known with the ELMAN neural network: under the roll channel fault effects, the output of jaw channel infrared earth instrument sensor is influenced to depart from 0 gradually.Yet the ELMAN neural network still can generate stable corresponding fault diagnosis result signal, thereby has proved the validity of ELMAN neural network under output coupling situation.

C. time-domain signal diagnosis generalization contrast

In order to verify the ELMAN neural network in the superiority aspect the live signal diagnosis generalization, with this paper ELMAN method and traditional BP NN method at infrared earth instrument roll channel output shut fault data and carried out the time-domain signal diagnosis performance and contrast.By Figure 12 and Figure 13 as can be seen, when the more steady 1500～1515s of input signal (being that signal only contains steady state characteristic), back-propagating neural network (BPNN) network and the output of ELMAN network all can keep stablizing.Yet when input signal occurs fluctuating 1520-1540s among a small circle when (temporal signatures appears in signal), BPNN diagnosis output hopping scope is very big, and mal-condition down even surpassed detection threshold.And the ELMAN neural network still can maintain in the threshold range, thereby has proved that the ELMAN neural network diagnoses superiority aspect extensive at time-domain signal.

D. network convergence contrast

Adopt the superiority of improvement gradient learning algorithm in order to contrast this paper in ELMAN neural network convergence index, this paper is respectively at adopting the two kinds of ELMAN neural network failure isolation module training constringency performances (MSE mean square deviation) that improve gradient descent algorithm and do not adopt gradient descent algorithm to contrast convergence curve such as Figure 14 and shown in Figure 15.Not improving the ELMAN neural network compares and improves the ELMAN neural network slow aspect rate of convergence (not improving for 2500 generations just restrains, and just convergence of 2000 generations after improving), and also (not improving the convergence mean square deviation is 0.00187 not as good as improving the ELMAN neural network aspect convergence precision, and improve the convergence mean square deviation is 0.00098), thus verified and improved the validity of ELMAN aspect the network training convergence.

Claims (4)

1. the blind system fault that is used for the spacecraft real time signal processing detects and partition method, may further comprise the steps:

(1) online detection:

(1.1) with i=k-N ..., k detects spacecraft diagnosis basis live signal p constantly _iInput fault detects neural network a ₂(i)=f ₂(LW _2,1a ₁(i)+b ₂), a wherein ₁(i)=f ₁(IW _1,1p _i+ LW _1,1a ₁(i-1)+b ₁), IW _1,1, LW _1,1, b ₁, LW _2,1, b ₂Be the structural parameters of fault detect neural network, f ₁() is the hidden layer transition function of fault detect neural network, f ₂() is the output layer transition function of fault detect neural network, a ₂(i) be one-dimensional vector, a ₁(i) be the h dimensional vector, h 〉=(m*1)/2, m is diagnosis basis live signal p _iNumber, k be current detection constantly, N is a data window length;

(1.2) to i=k-N ..., k detects fault detect neural network output a constantly ₂(i) make Filtering Processing, obtain the filtering result $O\left(k\right)=\frac{1}{N+1}\underset{i=k-N}{\overset{i=k}{\mathrm{\&Sigma;}}}{a_2}^2\left(i\right);$

(1.3) if filtering as a result O (k) smaller or equal to detection threshold ε ₀, then system is normal, otherwise the system failure enters step (2);

(2) online isolation:

(2.1) with i=k-N ..., k detects spacecraft diagnosis basis live signal p constantly _iInput fault is isolated neural network ${\tilde{a}}_2\left(i\right)={\tilde{f}}_2(\tilde{L}{W}_{\mathrm{2,1}}{\tilde{a}}_1\left(i\right)+{\tilde{b}}_2),$ Wherein ${\tilde{a}}_1\left(i\right)={\tilde{f}}_1(\tilde{I}{W}_{\mathrm{1,1}}{p}_i+\tilde{L}{W}_{\mathrm{1,1}}{\tilde{a}}_1(i-1)+{\tilde{b}}_1),$ $\tilde{I}{W}_{\mathrm{1,1}},\tilde{L}{W}_{\mathrm{2,1}},{\tilde{b}}_1,\tilde{L}{W}_{\mathrm{2,1}},{\tilde{b}}_2$ Be the structural parameters of fault isolation neural network,

Be the hidden layer transition function of fault isolation neural network,

Be the output layer transition function of fault isolation neural network,

For

Dimensional vector, $\tilde{n}\&GreaterEqual;{\log }_2\mathrm{FN},$ FN is the fault sum,

For

Dimensional vector, $\tilde{h}\&GreaterEqual;(m*\tilde{n})/2;$

(2.2) to i=k-N ..., k detects fault isolation neural network output constantly

Make Filtering Processing, obtain the filtering result $\tilde{O}\left(k\right)={\left[\begin{array}{ccc}{\tilde{O}}_1& ...& {\tilde{O}}_{\tilde{n}}\end{array}\right]}^T=\frac{1}{N+1}\underset{i=k-N}{\overset{i=k}{\mathrm{\&Sigma;}}}{{\tilde{a}}_2}^2\left(i\right),$ T represents transposition;

(2.3) the filtering result that step (2.2) is obtained

In each the dimension element

Respectively with isolation threshold Make comparisons, smaller or equal to zero, then the logic output valve of this dimension element correspondence is H0 as if comparative result, otherwise the logic output valve of this dimension element correspondence is H1, thereby obtains the filtering result vector

Corresponding logic output sequence;

(2.4) in the fault mode matching list, mate fault mode according to the logic output sequence.

2. blind system fault according to claim 1 detects and partition method, it is characterized in that the structural parameters IW of described fault detect neural network _1,1, LW _2,1, b ₁, LW _2,1, b ₂Determine in the following manner:

A1, make X (t)=[IW _1,1(t) LW _2,1(t) b ₁(t) LW _2,1(t) b ₂(t)] ^T, initialization X (1), X (2), t=3;

A2, calculating $\left\{\begin{array}{c}X\left(t\right)=X(t-1)+\mathrm{\&Delta;X}(t-1)\\ \mathrm{\&Delta;X}(t-1)=\mathrm{\&alpha;\&Delta;X}(t-2)+\mathrm{\&alpha;}{l}_r\frac{d_{\mathrm{perf}}}{\mathrm{dX}}\end{array}\right.,$ Wherein, detect neural network performance perf and the derivative that detects weight and migration parameter X (t) $\frac{d_{\mathrm{perf}}}{\mathrm{dX}\left(t\right)}=\frac{d}{\mathrm{dX}\left(t\right)}(\underset{s=1}{\overset{G}{\mathrm{\&Sigma;}}}(a_2\left(s\right)-y\left(s\right))\&CenterDot;{(a_2\left(s\right)-y\left(s\right))}^T),$ Y (s) is the desired output of s fault detect neural network during the moment, and G is the time series total length of normal and fault training sample, l _rThe learning rate of expression fault detect neural network, α represents the momentum term of fault detect neural network;

A3, if detect neural network performance perf ∈ [0,0.001], the structural parameters equivalence of then fault detect neural network finish, otherwise t=t+1 returns steps A 2 in X (t).

3. blind system fault according to claim 1 detects and partition method, it is characterized in that the structural parameters of described fault isolation neural network $\tilde{I}{W}_{\mathrm{1,1}},\tilde{L}{W}_{\mathrm{2,1}},{\tilde{b}}_1,\tilde{L}{W}_{\mathrm{2,1}},{\tilde{b}}_2$ Determine in the following manner:

B1, order $\tilde{X}\left(t\right)={\left[\begin{array}{ccccc}\tilde{I}{W}_{\mathrm{1,1}}\left(t\right)& \tilde{L}{W}_{\mathrm{2,1}}\left(t\right)& {\tilde{b}}_1\left(t\right)& \tilde{L}{W}_{\mathrm{2,1}}\left(t\right)& {\tilde{b}}_2\left(t\right)\end{array}\right]}^T,$ Initialization

T=3;

B2, calculating $\left\{\begin{array}{c}\tilde{X}\left(t\right)=\tilde{X}(t-1)+\mathrm{\&Delta;}\tilde{X}(t-1)\\ \mathrm{\&Delta;}\tilde{X}(t-1)=g\left(\mathrm{\&Delta;}\tilde{X}(t-2)\right)\mathrm{\&Delta;}\tilde{X}(t-2)+\widetilde{\mathrm{\&alpha;}}{\tilde{l}}_r\frac{d_{p\tilde{e}\mathrm{rf}}}{d\tilde{X}\left(t\right)}\end{array}\right.\text{,}$ Wherein, isolation network performance

With isolation weight and migration parameter

Derivative $\frac{d_{p\tilde{e}\mathrm{rf}}}{d\tilde{X}\left(t\right)}=\frac{d}{d\tilde{X}\left(t\right)}(\underset{\tilde{s}=1}{\overset{\tilde{G}}{\mathrm{\&Sigma;}}}({\tilde{a}}_2\left(\tilde{s}\right)-\tilde{y}\left(\tilde{s}\right))\&CenterDot;{({\tilde{a}}_2\left(\tilde{s}\right)-\tilde{y}\left(\tilde{s}\right))}^T),$

For

The desired output of fault isolation neural network in the time of constantly,

Be the length of time series of fault training sample, The learning rate of expression fault isolation neural network,

The momentum term of expression fault isolation neural network, $g\left(\mathrm{\&Delta;}\tilde{X}(t-2)\right)=\mathrm{\&Delta;}\tilde{X}(t-2)\&CenterDot;\exp (-\mathrm{\&Delta;}{\tilde{X}}^2(t-2));$

B3, if isolate the neural network performance $p\tilde{e}\mathrm{rf}\&Element;\left[\mathrm{0,0.001}\right],$ Then the structural parameters equivalence of fault isolation neural network in

Finish, otherwise t=t+1 returns step B2.

4. blind system fault according to claim 1 detects and partition method, it is characterized in that described transition function f ₁(), f ₂(),

In employing tansig (), logsig (), the purelin () function any one.

Images